DESCRIPCIÓN Y VALORACIÓN DEL GRADO DE CONSERVACIÓN DE LAS ESPECIES DE INTERÉS COMUNITARIO COMUNITARIO

6.5.1 From Sentiment to Intent Analysis. Discourse and different pragmatic context can en-

hance sentiment analysis systems. However, knowing what a holder likes and dislikes is only a first step in the decision making process. Consider the statements in Examples (24), (25), (26), and (27):

(24) I don’t like Apple’s policy overall, and will never own any Mac products. (25) I wish to buy a beautiful house with a swimming pool.

(26) How big is the screen on the Apple iPhone 4S? (27) I am giving birth in a month.

If we look at these examples from a sentiment analysis point of view, only the first sentence, in Example (24), would be classified as negative, the other examples being objective.9 However, in addition to a negative opinion, the writer in Example (24) explicitly states their intention not to buy Mac products, which is not good news for Apple. In Example (25), the writer wishes for a change in their existing situation, but there is no guarantee that this wish will lead to forming an intention to buy a new house in the future. In Example (26), the writer wants to know about others’ opinions and, based on these opinions, they may or may not be inclined to buy an iPhone. Finally, in Example (27), one can infer that the writer may want to buy baby products that may help Web sites to provide the most appropriate ads to display. These last two examples are typical of implicit intentions.

Knowing about the holder’s future actions or plans from texts is crucial for decision makers: Does the writer intend to stop using a service after a negative experience? Do they desire to purchase a product or service? Do they prefer buying one product over another? Intent analysis attempts to answer these questions, focusing on the detection of future states of affairs that a holder wants to achieve.

We use the term intent as a broader term that covers desires, preferences, and in-

tentions, which are mental attitudes contributing to the rational behavior of an agent.

These attitudes play a motivational role and it is in concert with beliefs that they can move us to act (Bratman 1990). Indeed, before deciding to perform an action, an agent considers various desires, which are states of affairs that the agent, in an ideal world, would wish to be brought about. Desires may be in conflict and are thus subject to inconsistencies. Among these desires, only some can be potentially satisfied. The chosen desires that the agent has committed to achieving are called intentions (Bratman 1990; Wooldridge 2000; Perugini and Bagozzi 2004). Intentions cannot conflict with each other and have to be consistent. This constitutes an important difference between desires and intentions. This distinction has been formalized in the Belief-Desire-Intention model (Bratman 1990), an intention-based theory of practical reasoning, namely, reasoning directed toward actions.

Desires may be ordered according to preferences. A preference is commonly defined as an asymmetric, transitive ordering by an agent over outcomes, which are understood as actions that the agent can perform or goal states that are the direct result of an action of the agent. For instance, an agent’s preferences may be defined over actions like buy a new car or by its end result like have a new car. Among these outcomes, some are acceptable for the agent (i.e., the agent is ready to act in such a way as to

9 A standard system that does not account for the volitive modality of wish would also classify Example (25) as positive.

realize them) and some outcomes are not. Among the acceptable outcomes, the agent will typically prefer some to others. Preferences are not opinions. Whereas opinions are defined as a point of view, a belief, a sentiment, or a judgment that an agent may

have about an object or a person, preferences involve an ordering on behalf of an agent

and thus are relational and comparative. Opinions concern absolute judgments towards objects or persons (positive, negative, or neutral), and preferences concern relative judg- ments towards actions (preferring them or not over others). The following examples illustrate this.

(28) The movie is not bad.

(29) The script for the first season is better than the second one. (30) I would like to go to the cinema. Let’s go and see Madagascar 3.

Example (28) expresses a direct positive opinion towards the movie, but we do not know if this movie is the most preferred. Example (29) expresses a comparative opinion about two movies with respect to their shared features (script). If actions involving these movies (e.g., seeing them) are clear in the context, such a comparative opinion will imply a preference, an ordering the first season scenario over the second. Finally, Example (30) expresses two preferences, one depending on the other. The first is that the speaker prefers to go to the cinema over other alternative actions; the second is: Given the option of going to the cinema, they want to see Madagascar 3 over other possible movies.

Reasoning about preferences is also distinct from reasoning about opinions. An agent’s preferences determine an order over outcomes that predicts how the agent, if they are rational, will act. This is not true for opinions. Opinions have at best an indirect link to action: I may not absolutely love what I am doing right now, but do it anyway because I prefer that outcome to any of the alternatives.

6.5.2 Intent Detection: Main Approaches. Acquiring, modeling, and reasoning with desires,

preferences, and intentions are well-established fields in artificial intelligence (Cohen and Levesque 1990; Georgeff et al. 1999; Brafman and Domshlak 2009; Kaci 2011). Predicting user intentions from search queries and/or the user’s click behavior has also been extensively studied in the Web search community to assist the user to search what they want more efficiently (Chen et al. 2002; Wang and Zhang 2013). There is, however, little research that investigates how to extract desires, preferences, and intentions from users’ linguistic actions using NLP techniques. We survey here some existing work.

Desire extraction. Wish and desire detection from text have been explored by Goldberg

et al. (2009). They define a wish as “a desire or hope for something to happen” and propose an unsupervised approach that learns if a given sentence is a wish or not. Given that the expression of wishes is domain-dependent, they first exploit redundancy in how wishes are expressed to automatically discover wish templates from a source domain. These templates are then used to predict wishes in two target domains: product reviews and political discussions. The source domain is a subset of the WISH corpus composed of about 100,000 multilingual wish sentences collected over a period of 10 days in December 2007, when Web users sent in their wishes for the new year. Peace on

earth, To be financially stable, and I wish for health and happiness for my family, are typical

sentences. Extraction suggestions for products using templates has also been explored for tweets (Dong et al. 2013). Using a small set of hand-crafted rules, Ramanand et al. (2010) focus on two specific kinds of wishes characteristic of product reviews: sentences

that make suggestions about existing products, and sentences that indicate the writer is interested in purchasing a product. The same approach has been used in Brun and Hag`ege (2013) to improve feature-based sentiment analysis of product reviews. It is, however, limited, since the system only detects those wishes that match previously defined rules.

Preference extraction. Preference extraction from text has been investigated with the study

of comparative opinions (Jindal and Liu 2006a, 2006b; Ganapathibhotla and Liu 2008; Yang and Ko 2011; Li et al. 2013a). Given a comparison within a sentence, this task involves two steps. First extract entities, comparative words, and entity features that are being compared; then, identify the most preferred entity. In Example (29), the first

season and second season are the entities, better than the comparative, script the entity

feature, and the first season the preferred entity. This approach is quite limited, because it either only focuses on the task of identifying comparative sentences without extracting the comparative relations within the sentences, or when it does, it only considers com- parisons at the sentence level, even sometimes with the assumption that there is only one comparative relation in a sentence. However, for reasoning with preferences, it is unavoidable to consider more complex comparisons with more than one dependency at a time and with a higher level than just the sentence, in order to manage all the preference complexity. Cadilhac et al. (2012) explore such an approach to automatically extract the preferences and their dependencies within each dialogue move in negotia- tion dialogues. They perform the extraction in two steps: first the set of outcomes; then, how these outcomes are ordered. Those extracted preferences are then used to predict trades in the win–lose game Settlers of Catan (Cadilhac et al. 2013).

Intention extraction. As for desires and preferences, intention extraction is also formu-

lated as a classification problem: deciding whether a sentence expresses an intention or not. Sujay and Yalamanchi (2012) focus on explicit intentions and propose to categorize text according to the type of intentions it expresses among wish, praise, complain, buy, and so on. Using a naive bag-of-words approach, they achieve an accuracy of almost 67% on a social media corpus. Chen et al. (2013) also focus on explicit intentions in discussion forums such as I am looking for a brand new car to replace my old Ford Focus. The authors observe that this classification problem suffers from noisy data (only a few sentences express intentions) and domain-dependency of features indicating the negative class (i.e., non-intention). To deal with these issues, Chen et al. propose a transfer learning method that first classifies sentences using labeled data from a given source domain, and then applies the classifier to classify the target unlabeled data. Transfer learning has also been applied to detect implicit intentions in tweets following a two-step procedure (Ding et al. 2015): First, determine whether the sentence involves a consumption intention. If it does, extract intention words.

In summary, we see intent analysis as orthogonal and supplementary to sentiment analysis, which focuses on past/present holder’s states. This is why we believe that intent detection would benefit from being built on top of sentiment analysis systems, since positive or negative sentiments are often expressed prior to future actions. 7. When Linguistics Meets Computational Linguistics: Future Directions

We firmly believe that future developments in sentiment analysis need to be grounded in linguistic knowledge (and also extra-linguistic information). In particular, dis- course and pragmatic phenomena play such an important role in the interpretation of

evaluative language that they need to be taken into account if our goal is to accurately capture sentiment. The dynamic definition of sentiment that we have presented in- cludes update functions that allow for different contextual aspects to be incorporated into the calculation of sentiment for evaluative words and expressions, and can be applied at all levels of language. We see the use of linguistic and statistical methods not as mutually exclusive, but as contributing to each other. For instance, rather than general n-gram bag-of-words features, other features from discourse can be used to train classifiers for sentiment analysis. Contextual features can be deployed to detect implicit evaluation, and to accurately capture the meaning in figurative expressions.

We showed in this survey that including discourse information into opinion analy- sis is definitively beneficial. Discourse has also been successfully deployed in machine translation (Hardmeier 2013), natural language generation (Ashar and Indukhya 2010), and language technology in general (Taboada and Mann 2006a; Webber et al. 2012). Incorporating discourse into sentiment analysis can be done by relying either on shallow

discourse processing (using specific discourse markers, leveraging the notion of topicality,

zoning, and social network structure), or through full discourse parsing, exploiting the entire discourse structure of a document. The shallow approach has been shown to be effective when experimented on movie/product review data, and there is an increasing amount of work on other kinds of data, such as blogs (Liu et al. 2010; Chenlo et al. 2013) and tweets (Mukherjee and Bhattacharyya 2012), where links across posts and the stream of related posts are investigated. The effectiveness of full discourse, however, strongly depends on the availability of powerful tools, such as discourse parsers. Com- pared with syntactic parsing and shallow semantic analysis, discourse parsing is not as mature. To date, the performance of parsers is still considerably inferior compared with the human gold standard, although significant advances have been made in the last few years (Muller et al. 2012; Ji and Eisenstein 2014; Feng 2015; Joty et al. 2015; Surdeanu et al. 2015; Perret et al. 2016), and we expect improvements to continue. Automatic discourse segmentation has attained high accuracy. For example, Fisher and Roark (2007) report an F-score of 90.5% in English. Discourse relations remain nonetheless hard to detect, due in part to the ambiguity of discourse markers, and to implicit relations. End-to-end parsing involving structured prediction methods from machine learning is also still in development. For example, Ji and Eisenstein (2014) report an accuracy of 60% for discourse relation detection and Joty et al. (2015) achieve above 55% for text-level relation detection in the RST Treebank. Muller et al. (2012) also achieve between 47% and 66% accuracy on the ANNODIS corpus, annotated following SDRT. This may explain why most state-of-the-art NLP applications that rely on discourse do not yet offer a substantial boost compared with discourse-unaware systems.

An additional problem is the domain dependence of many of the existing parsers, which have been trained on newspaper articles, mostly versions of the Penn corpus of Wall Street Journal articles, either in its RST annotation (Carlson et al. 2002), SDRT annotation (Afantenos et al. 2012), or the Penn Discourse TreeBank annotation (Prasad et al. 2008). It is no surprise, then, that they do not perform very well on reviews. A possible solution would be to train a parser on gold discourse structure annotations and sentiment labels, such as the SFU Review Corpus (Taboada et al. 2008) or the CASOAR Corpus (Benamara et al. 2016). On the negative side, such corpora are too small to train a discourse parser and a competitive sentiment analysis system. On the positive side, review-style documents are relatively short , which can make parsers less sensitive to errors due to long dependency attachments. Also, given that not all discourse relations are sentiment relevant, the number of relations to be predicted can be reduced. This might concern framing relations like BACKGROUND and CIRCUMSTANCE but also

some temporal relations such as SEQUENCE or SYNCHRONOUS. On the other hand, relations can be grouped according to their similar effects on both subjectivity and polarity analysis. One possible grouping could be argumentative relations that are used to support (e.g., MOTIVATION, JUSTIFICATION, INTERPRETATION) or oppose (e.g., CONTRAST, CONCESSION, ANTITHESIS) claims and theses, causal relations (e.g., RESULT, CONDITION), and structural relations (e.g., ALTERNATIVE, CONTINUATION). Thematic

relations like ELABORATION, SUMMARY, and RESTATEMENThave also a strong impact on evaluative discourse. Their effect is, however, close to support relations. We believe that discarding certain relations and grouping others will make discourse parsers more reliable.

Besides domain dependency, parsing is by definition theory-dependent, which means that a system trained to learn RST relations fails to predict SDRT or PDTB relations. Indeed, each theory has its own hierarchy of discourse relations, but relations tend to overlap or be related in a few specific ways: A relation R in one approach can correspond to several relations in another approach and vice versa; a relation may be defined in one approach but not taken into account in another; and, finally, relations across approaches may have similar names, but different definitions. One solution to this problem is to map relations across approaches to a unified hierarchy. Merging different discourse relation taxonomies has several advantages. First of all, for classification tasks such as discourse parsing, access to larger amounts of data is likely to yield better results. Secondly, and from a more theoretical point of view, we think that differences across approaches are minimal, and a unified set of relations is possible. Third, a unified set of discourse relations would allow us to compile a list of discourse markers and other signals for those relations, which would also benefit discourse annotation. Recent efforts to merge existing discourse relation taxonomies and annotations should help improve discourse parsing (Benamara and Taboada 2015; Rehbein et al. 2015). Notable also is work being carried out within the COST Action TextLink, a pan-European initiative to unify definitions of relations and their signals across languages (http://www.textlink.ii.metu.edu.tr/).

Once powerful discourse parsers are developed, the argumentative structure of evaluative text can be fully exploited. Processing arguments for sentiment analysis is still at an early stage and we feel that recent progress in argument mining will likely spur new research in this direction (Bex et al. 2013; Stab and Gurevych 2014; Peldszus and Stede 2015).

We see sentiment analysis not as an aim per se but as a first step in processing and understanding large amounts of data. Indeed, sentiment analysis has strong interac- tions with social media (Farzindar and Inkpen 2015), big data (Arora and Malik 2015), and, more importantly, with modeling human behavior, that is, how sentiment trans- lates into action. We defined “the sentiment to action” process as intent detection (cf. Section 6.5), an area which gives linguistic objects a predictive power such as predicting voter behavior and election results (Yano et al. 2013; Qiu et al. 2015), predicting decep- tion (Fitzpatrick et al. 2015), or intention to buy (Ding et al. 2015). Predictions can also be derived on the basis of extra-linguistic sources of information such as characteristics of the author and their online interactions (Qiu et al. 2015). State-of-the-art approaches are still heavily dependent on bag-of-words representations. We believe that predicting a user’s future actions from text (and speech) needs to integrate models from artificial intelligence with NLP techniques to find specific intent signals, such as changes in the argumentation chain; the social relationship between discourse participants; topic changes; user’s beliefs; the sudden use of sentiments or emotions of a certain type (like aggressive expressions); or the correlation between genre and the use of specific

linguistic devices. Intent detection is an emerging research area with great potential in business applications (Wang et al. 2015).

In summary, we believe that the discourse turn that computational linguistics is experiencing can be successfully combined with data-driven methods as part of the effort to accurately capture sentiment and evaluative language.

Acknowledgments

This research was funded by a Discovery Grant from the Natural Sciences and Engineering Council of Canada to Maite Taboada, and ERC grant 269427 (STAC). We thank Nicholas Asher, Paola Merlo, and the two anonymous reviewers for their comments or suggestions. We take

responsibility for any errors that may remain.

References

Abbasi, Ahmed, Hsinchun Chen, and Arab Salem. 2008. Sentiment analysis in multiple languages: Feature selection for opinion classification in Web forums. ACM

Transactions on Information Systems,

26(3):1–34.

Afantenos, Stergos, Nicholas Asher, Farah Benamara, Myriam Bras, C´ecile Fabre, Mai Ho-Dac, Anne Le Draoulec, Philippe Muller, Marie-Paule P´ery-Woodley, Laurent Pr´evot, Josette Rebeyrolles, Ludovic Tanguy, Marianne Vergez-Couret, and Laure Vieu. 2012. An empirical resource for discovering cognitive principles of discourse organisation: The ANNODIS corpus. In Proceedings of the

Eighth International Conference on Language Resources and Evaluation, LREC 2012,

pages 2727–2734, Istanbul.

Agarwal, Shashank and Hong Yu. 2009. Automatically classifying sentences in full-text biomedical articles into introduction, methods, results and discussion. Bioinformatics,

25(23):3174–3180.

Agarwal, Apoorv, Fadi Biadsy, and